A Review of Skin Cancer Classifications and

Management Strategies
line 1: 1st Given Name Surname line 1: 2nd Given Name Surname line 1: 3rd Given Name Surname
line 2: dept. name of organization line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 3: name of organization line 3: name of organization line 3: name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 4: City, Country line 4: City, Country line 4: City, Country
line 5: email address or ORCID line 5: email address or ORCID line 5: email address or ORCID

line 1: 4th Given Name Surname line 1: 5th Given Name Surname line 1: 6th Given Name Surname
line 2: dept. name of organization line 2: dept. name of organization line 2: dept. name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 3: name of organization line 3: name of organization line 3: name of organization
(of Affiliation) (of Affiliation) (of Affiliation)
line 4: City, Country line 4: City, Country line 4: City, Country
line 5: email address or ORCID line 5: email address or ORCID line 5: email address or ORCID

Abstract—Three different kinds of skin cancer can be extensive knowledge about the molecular profile and
discovered on Earth: Cutaneous malignancies include essential risk indicators associated with these pathological
melanoma, basal cell carcinoma and squamous cell carcinoma conditions.
of skin. The various types of skin cancer have their causes,
effects, or outcomes and the approaches of handling the II. LITERATURE SURVEY
treatments separately. Out of all types of cancer, skin cancer is
Dildar et al. (2021) identify skin diseases, particularly
the most widespread type of tumor. Consequently, the aim of
this review is to present a detailed study on skin cancer
skin cancer from genetic mutations as highly perilous. Early
subtypes by reviewing the various types of skin cancer, their detection, crucial due to high fatality and treatment costs is
molecular characteristics, risk factors and clinical possible through various methods. This research
manifestations. Moreover, it explores the available modern systematically reviews deep learning algorithms for early
diagnostic procedures like molecular testing, dermoscopy, and detection, examining high-quality studies and utilizing tools,
histopathology that facilitated better staging of cancers and graphs, tables, methodologies, and frameworks. Schierbeck
better early diagnosis. The study also explores the change of et al. (2019) report 15,000 annual skin cancer cases in
management strategies currently in use such as radiation Denmark (population 5.7 million), caused mainly by UVR,
therapy, surgical excisions, immunotherapies, photodynamic immunosuppressive treatment, and irradiation. Daghrir, Tlig,
therapy, targeted therapies among others. It is therefore et al. (2020) propose a a combined method for melanoma
genetic screening and more so the personalized medicine which detection, integrating machine learning and deep learning. A
are right at the fore front of the treatment options for the CNN and two traditional classifiers based on lesion
advanced melanoma cases. Health promotion practices, characteristics are combined using a majority voting system,
including community awareness and sun protection measures enhancing accuracy and aiding early detection.
are other topics of debate. The presentation of a concept
focusing on the involvement of a plurality of disciplines in skin Kadampur and Al Riyaee (2020) identify melanoma,
cancer patients care is one of the goals of this research, and basal cell carcinoma, and squamous cell carcinoma as
thus contributing to better results in the fate of patients with prevalent skin cancers, each with unique challenges. This
this illness. To accomplish this goal, both the old and the new study reviews these cancers' molecular mechanisms, risk
approach will be looked at both ways. factors, and symptoms. It highlights diagnostic advancements
like molecular testing and dermoscopy, and explores
Keywords— Skin cancer, Melanoma, Basal cell carcinoma melanoma management strategies including genetic profiling
(BCC), Squamous cell carcinoma (SCC), Clinical symptoms,
and tailored treatments. Public health programs and sun
Diagnostic methods, Dermoscopy, Molecular testing, Skin cancer
staging, Skin cancer prevention
protection are emphasized for prevention, advocating a
multidisciplinary approach to enhance therapy and patient
I. INTRODUCTION outcomes. The work of Berhil, Benlahmar, & Labani,
(2020), reveals how human capital analysis with the help of
The three commonly occurring skin cancer types in the
HR analytics positively affects profitability of companies.
world are melanocytic carcinoma, basalioma, and squamous
This review provides a comprehensive analysis of HR risks
cell carcinoma. A primary type of skin cancer is melanoma,
and challenges and the latest AI advancements addressing
and most skin cancer incidents are melanoma. Each of these
these issues. Summarizing IT solutions from 2008 to 2018, it
subtypes presents certain difficulties in tracing the
serves as a reference for computer scientists, outlining how
development of the disease, its prognosis, and therapy.
AI can transform HR processes and decision-making..
Squamous cell carcinoma and basal cell carcinoma are less
dangerous and most frequently occurring compared to the Kliegr, Bahník, & Fürnkranz (2021) emphasize that
melanoma that is known to be highly invasive and fatal. The understanding machine learning interpretability of models
fact is that to enhance the diagnosis, prediction, and therapy require delving into cognitive science beyond syntactic
of numerous forms of skin cancer it is critical to have comprehension. They discuss twenty cognitive biases

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

affecting interpretability and provide debiasing solutions, resilience, and performance. While deep learning shows
calling for empirical studies to bridge the gap between promise, further research and innovation are needed to
cognitive psychology and machine learning. Hajiarbabi overcome existing challenges.
(2023) introduces a technique for differentiating melanoma
Thrall, et al. (2018), the ongoing development in
from benign skin abnormalities using preprocessing steps,
computing power, large datasets, and deep learning
image augmentation, and a CNN. A three-stage analysis
algorithms have sparked interest in AI in imaging. AI
focuses on lesion centres, with results combined via a fully
presents both opportunities and risks, requiring
connected neural network. The method achieves 88.5%
considerations of nomenclature, data sharing, and validation
recall, 91.75% precision, 94.42% accuracy, and an AUC of
across imaging platforms and patients. AI can help
0.94, demonstrating effectiveness.
radiologists prioritize cases and enhance diagnostic results.
Lu & Firoozeh (2022) highlight the significant issue of Though some fear AI may replace radiologists, this paper
skin cancer, 40% of all cancers, with over 300,000 new argues that AI will complement human efforts for better and
melanoma cases in 2018. Traditional treatments like surgery, faster diagnoses. Current limitations include lack of
radiation, and chemotherapy have been the mainstay of knowledge and computer resources, addressable by technical
cancer care for decades, but new image processing support. AI's impact on imaging will be measured by patient
technologies show promise for early detection and treatment. health, time saved, improved diagnostic precision, and more
An enhanced XceptionNet model, with improved free time for radiologists. Radiologists are expected to lead
performance metrics, demonstrates that modern machine AI integration in medical imaging.
learning methods can significantly enhance diagnosis
Dildar et al., 2021 DNA damage to skin cells causes
accuracy. In the work of Khanzode & Sarode (2020), the
mutations, allowing the disease to spread. Early detection is
history of ET is described from the concept of traditional
crucial for better treatment outcomes. Skin cancer is rising,
form of AI to the integration of AI in the creation of ICAI
with high fatality and treatment costs, emphasizing early
systems. The paper synthesizes literature to explore
detection. Researchers have developed techniques to
applications of AI in education, emphasizing its potential for
differentiate melanoma from benign disorders, focusing on
academic professionals and researchers.
lesion symmetry, colour, size, and shape. Connor, C. W.
According to Ullah, et al. (2020), smart cities aim to (2019), the purpose of this inquiry is to examine the AI and
tackle urbanization, energy management, environmental ML advancements and their growing use in commerce, such
sustainability, and quality of life using ICT. AI, ML, and as image recognition, language translation, and speech. AI
DRL are emerging as crucial in addressing these complex faces unique challenges in risk-sensitive domains like
issues. Key applications include energy-saving smart grids, anaesthesiology due to high reliability and complex decision-
intelligent transportation systems, cybersecurity, and making. The review introduces AI and ML to
improved 5G communication services using UAVs. The role anaesthesiologists, examines algorithm decision-making, and
of these technologies in smart healthcare systems is also explores clinical applications, suggesting a new paradigm of
examined. The report identifies research roadblocks and ML-driven discovery in anaesthesiology.
suggests approaches to enhance smart city initiatives with AI,
Wysocki, et al. (2023), this paper assesses clinically
ML, and DRL. Hermosilla, et al. (2024), reviews computer-
relevant explainable Clinical Decision Support Machine
assisted technology for early skin cancer detection via
Learning models in healthcare. It captures the dynamics of
dermatoscopic image processing. It analyses 45 studies,
explanation models trained with ML algorithms in clinical
focusing on algorithms, accuracy, and validation. Despite
settings. Results show these approaches require significant
advancements, challenges like image quality and reliance on
effort, are prone to logical fallacies, and heavily depend on
human interpretation persist. Further research is needed to
models. While explanations boost confidence and safety,
improve practicality and reliability of these technologies,
they also reveal model flaws. The study identifies advantages
paving the way for future advancements in skin cancer
of explainable ML models, like eliminating bias, aiding
diagnostics.
clinical decisions, and helping less experienced healthcare
Lemaignan, et al. (2017), this study identifies personnel find information. It calls for rethinking ML
cognitive and decisional challenges for effective human- explanations to enhance therapeutic decision-making and
robot collaboration. Essential cognitive abilities for robots learning.
include geometric reasoning, multi-perspective evaluation,
As the authors Rupali M and Amit P pointed out in this
and affordance analysis. Knowledge models accounting for
paper, this paper reviews the history of AI and ML and their
human-robot differences, multi-modal communication, and
impact on employment, contrasting modern business and
awareness-based task planning are crucial. The paper
daily life with manual past practices. AI and ML are
emphasizes explicit knowledge management for natural
revolutionizing technologies that minimize human
human-robot interactions and calls for incorporating human-
intervention and increase efficiency. The paper highlights
level meanings into AI decisions. It highlights the importance
how these technologies, once criticized, have ushered in an
of robust knowledge management for successful human-
era of high efficiency and accuracy, and explores their
robot collaboration.
impact on the future.
Wu, et al. (2022), focus on deep learning classification
From Collenette, et al. (2023), the current study
methods for skin cancer diagnosis. Challenges include data
demonstrates the importance of explainability in AI tools for
asymmetry, limited accessibility, and model adaptation. It
legal thinking, making solutions easily consumable for end
starts with open-source datasets and common dermatological
users. Experimental evaluation showed 97% accuracy in real-
images for diagnosis, then examines convolutional neural
world scenarios, with user polls indicating high
networks in skin cancer classification. The study addresses
understandability. AI supports fast and effective legal
issues like data disparities, domain flexibility, model
decisions, demonstrating its potential for transparency and by methods such as radar, infrared energy, computer-
reliability in complex legal domains. generated patterns, ultrasonic energy, and X-rays. The type
of camera sensor used depends on the lighting conditions.
D. A. Winkler, explores advancements in robotics,
‘Image acquisition’ refers to capturing an image, where light
automation, and AI for nanomaterial synthesis and analysis.
energy hitting the sensor is transformed into digital images.
While automation has progressed, nanomaterials still require
Sensors convert incoming illumination into measurable
further development for ML methods in data analysis. The
voltages. Charge-Coupled Devices (CCDs) and
study examines traditional and deep learning approaches to
Complementary Metal-Oxide Semiconductors (CMOS) are
nano safety and nanomaterial characterization, identifying
common types of image sensors. CCD sensors offer high-
challenges and suggesting improvements. AI and ML
resolution images with minimal noise, while CMOS sensors
advancements could significantly enhance safe nanomaterial
are more susceptible to noise but consume less power. Image
production. Shorfuzzaman et al.(2021) Malignant melanoma,
acquisition involves a two-dimensional array of cells that
a deadly form of skin cancer, results from the uncontrolled
convert light into electrons, which are then processed to form
proliferation of melanocyte cells. Automated detection from
a digital image.
dermoscopic images using deep Convolutional Neural
Networks (CNNs) is gaining popularity, though their ‘black- Tumpa et al.(2021) Adequate data is essential for
box’ nature limits clinical use. This study suggests an developing automated melanoma detection methods. This
interpretable CNN-based stacked ensemble architecture for study used dermoscopy images from the ISIC archive and
early melanoma detection, using multiple CNN sub-models PH2 datasets to classify melanoma. The ISIC dataset
in a stacking ensemble framework with transfer learning. A contains 3600 JPEG images (224×224 pixels) of cancerous
meta-learner combines sub-model predictions for the final and benign lesions for training and testing. The PH2 dataset
result. includes 200 BMP files (768×560 pixels), with 160 benign
and 40 malignant cases, created by the Pedro Hispano
Early melanoma detection can save lives and is
Hospital Dermatology Service in Portugal and the University
achievable through computer-aided approaches. This study
of Portland’s School of Medicine. The images were taken in
proposes an ensemble technique combining CNNs and visual
RGB format, and the datasets were combined for network
texture feature extraction for automatic melanoma detection.
training.
It uses VGG-19 and a proposed network model for image
categorization, with kernel Principal Component Analysis C. Skin Image Preprocessing
(kPCA) for feature dimension reduction. The technique was Studied in 2016 by Fioravante Oliveira and other
tested on ISIC 2016, ISIC 2019, and PH2 datasets. Diame et researchers. Effective detection and analysis of pigmented
al.(2021) Melanoma, though part of all skin cancers, skin lesions require pre-processing. Challenges in image
accounts for over 75% of skin cancer deaths in the U.S. segmentation include hairs, reflected light, shadows, skin
Researchers developed automated methods for early lines, and air bubbles. The blue channel accentuates lesion
detection using machine learning to segment skin lesions. regions, and the original RGB image can be used in vector
New deep network architectures have improved processing without converting to other colour spaces like
segmentation and diagnosis, with a comparison of methods CIE or HSV. Scalar techniques can be applied to each colour
highlighting their advantages and disadvantages. channel and combined or processed using vector data
Rojaramani et al.(2020) Melanoma, a lethal skin processing algorithms.
condition, can be malignant or non-malignant. Benign Artifacts from lighting changes can affect skin lesion
melanoma involves non-cancerous moles, while malignant segmentation. A quadratic function models shading effects,
melanoma, the worst type, causes bleeding sores and spreads reducing these effects when converting images back to RGB
slowly from melanocytes. Early detection is crucial for from HSV. Otsu's thresholding technique (Glaister et al.
effective treatment. Dermatology imaging technology aims to 2013) segments colour-corrected images. Novel approaches,
detect melanoma early, often using stable monitoring and such as Monte Carlo sampling, correct lighting variance in
diagnostic success. Biopsies, though painful and time- macroscopic images. The approach uses Otsu's histogram
consuming, are commonly used for diagnosis. Computerized bimodality metric to find the best weights for converting an
approaches offer more accurate and time-saving alternatives. RGB image to a grey-level image. Shade-of-grey methods
(Barata, et al. 2014) correct dermoscopy images based on
III. SKIN CANCER IMAGE CAPTURING METHODS
light source hue. Motif-based set theory may be used to
A. Image acquisition improve the appearance of skin blemishes.
Mishra et al.(2017) Image acquisition is the first step in Colour Morphological filtering enhances lesion areas by
video or image processing, allowing the camera to capture reducing noise and low-contrast boundaries. Hair removal in
and transform images into editable data. It involves three macroscopic and dermoscopy images is crucial as it
stages: an energy-focusing optical system, energy reflected influences lesion borders. Early methods identify hair
from the object, and a device measuring the energy used. removal stages, particularly for dense black hair, using
Various cameras serve different purposes: x-ray sensitive thresholds to create binary masks. Anisotropic diffusion and
cameras for x-ray images, infrared-sensitive cameras for median filters reduce hair visibility and enhance
infrared images, and visual spectrum cameras for everyday segmentation accuracy while retaining important lesion data.
photos. Image processing begins with capturing images using These filters effectively handle noisy images, and iterative
measurable energy, such as light or electromagnetic waves. application of anisotropic diffusion reduces high-frequency
B. Image Acquisition Model noise and enhances edges. With the help of the supplied
improvements, the new method is desired not only to
Creating an image involves an illumination source and enhance edges but also to denoise images without
objects that reflect or absorb energy. Light can be produced disregarding important edges. Anisotropic diffusion filter
iteration 150 was the final check that halted the smoothening questionable, or malignant. The widely accepted ABCD rule,
process. introduced by Stolz in 1985, includes asymmetry, border
irregularity, color variation, diameter beyond 6 millimeters,
D. Image Processing: Filters for Noise Reduction and Edge and evolution. Diagnosing early-stage melanoma is
Detection challenging due to minor symmetry in shape and color. The
Smoothing Filters are used for blurring and noise Menzies technique classifies moles based on one color and
reduction. Two types are Smoothing Linear Filters and symmetry distribution. Yacin Sikkandar et al. (2021) utilize
Smoothing Non-Linear Filters. An average filter determines the Inception v4 model for feature extraction from segmented
the pixel values' average in a region using a mask. A skin lesion images. Older Inception techniques divide
weighted average filter elevates specific pixels by repeating blocks into sub-networks stored in the storage
multiplying their values. Larger masks increase blurring. region. A simple filter count tweak in Inception models
Convolution uses more pixels to determine the average maintains network quality and improves computation and
result. Median filters, a type of non-linear filter, remove salt training rates. Tensor Flow-based Inception techniques
and pepper noise by sorting pixel values in a region, minimize tensor matrices for backpropagation. Inception-v4
calculating the median, and assigning it to the pixel. creates typical possibilities for blocks of various grid sizes, as
Sharpening Filters highlight transition in intensity. Types shown in Table I, comparing models used for feature
include The Laplacian (Second Derivative), The Gradient extraction.
(First Order Derivative), and Sobel Operators. These filters
show intensity changes around edges. The first derivative TABLE I. QUALITATIVE COMPARISON MODELS WITH WELL KNOWN
shows pixel intensity fluctuations, while the second MODELS USED FOR FEATURE EXTRACTION

derivative indicates changes in direction. Title Classes Features

A Benign and 1. The form is
The Laplacian — Second Derivative used in this methodological characterized by the area, the
procedure is defined as follows, approach to the Melanoma: 2 Classes level of its asymmetry,
classification of compactness, and diameter.
dermoscopy
images 2. Colour features include
colour space, colour moment,
(1) centroids, standard deviation,
mean and centroid diameter.

The Gradient — First Order Derivative is good for 3. The GLCM is one of the
properties of textures
spotting preprocessing flaws. In image processing, the first
derivatives are computed based on the gradient’s magnitude.
Computer Melanomas and 1.Color
The pace at which the gradient changes direction is expressed Aided Diagnosis of Mealanocytic: 2
by this magnitude. This filter removes all of the image’s Mealanocytic Classes 2.Texure
isotropic features. Lesions

Sobel operator — Using the Gradient gives a resulting Pattern Benign Color and texture properties
image is smoothed while the edges are enhanced by this classification of Melanocytic can be used to derive
dermoscopy lesion and multiscale texture features and
sharpening filter, which makes use of a coefficient. In the
images: a Melanoma. 2 colour symmetry.
middle, that it employs a weight value of 2. The total of all perceptually Classes
the masks is zero, as predicted for a derivative operator, as uniform
can be seen in the examples. model

Using multiple preprocessing filters in image processing Automating Benign, Dysplastic The size comprises of the
is common to enhance training datasets for computer vision melanoma lesion and bounding rectangle, area,
prediction and Melanoma: 3classes traverse, and ploar
and machine learning. Noise reduction filters are often detection measurements. How rounded
applied before masks to focus on specific areas of an image. it is and how small it can be
Banerjee et al. (2020) highlight that dermoscopy, a costly made. Colour defines the
diagnostic method for skin cancer, can be replaced with present colour space
Blum's 'tape dermatoscopy' technique, which is cost-effective respectively.
Different facets of the slope
without compromising image quality. This involves using a
transparent adhesive on the lesion after immersing the region Segmentation and Melanoma, Color features: color space
in immersion fluid. Yacin Sikkandar et al. (2021) describe a Classification of Bullae,Seborrheic with color moment and RGB
two-step pre-processing method for skin lesion images. The skin keratosis,Shingles histogram. Texture features:
lesions for disease and GLCM with Haralick features
initial step includes format conversion and area of interest diagnosis Squamous cell: 5
(ROI) recognition, followed by hair removal, as hair Classes
significantly impacts detection and classification. The image
is first converted to grayscale, and then morphological image A. Skin Lesion Segmentation
processing, specifically a top hat filter, is used to identify and
remove thick, black hair. Factors such as equivalent grey levels or hues can impact
the segmentation process, alongside similarity criteria. Skin
IV. SKIN CANCER – FEATURE EXTRACTION METHODS lesion identification in images can be achieved using
thresholding and region-based segmentation with similarity
Banerjee et al. (2020) emphasize the importance of early criteria. Many methods begin with scalar photos, converting
lesion diagnosis in fighting skin cancer, highlighting the original color image to scalar data, like greyscale, for
dermoscopy as a vital non-intrusive tool for melanoma more efficient processing. Segmentation methods for vector
detection. Melanocytic skin lesions can be benign, images rely on data from individual color channels in the
original image, which is computationally intensive and dsNet for better pixel space properties. Skin lesion networks,
requires specific colour spaces. Table II discusses the various such as those using encoder-decoder FCNs, dense blocks,
techniques used in the segmentation strategy. and CRFs (Adegun et al. 2020), enhance performance with
less model complexity, though border segmentation remains
TABLE II. SEGMENTATION METHODS AND ITS DIFFERENT challenging. Ashour et al. (2021) discuss segmentation
TECHNIQUES
techniques that divide images into areas of interest using
Segmentation Method Technique region-based and edge-based approaches. Techniques like
Region-based Statistical region merging
Particle Swarm Optimization (PSO), Ant Colony
Optimization (ACO), and Krill Herd methods enhance
Iterative stochastic region merging segmentation. Optimized Laplacian of Gaussian (LoG)
Region growing energy and parallel computing improve the Active Contour
(AC) method. Shahamatnia and Ebadzadeh utilized revised
Active contour-based Gradient vector flow PSO and the Bat algorithm to avoid local minima.
Region-based active contour algorithm
TABLE III. QUALITATIVE COMPARISON MODELS WITH WELL KNOWN
Active contour without edges MODELS USED FOR SEGMENTATION

Expectation-maximization level set Segmentation

Title Classes
Method
Adaptive snake
A methodological 2 Classes:
Level set approach to the Benign and JSEG algorithm for
classification of Melanoma segmentation
AI-based Evolutionary computation dermoscopy images
k-means clustering Computer Aided
2 Classes: Segmentation by
Diagnosis of
Melanomas and Morphological
Neural networks Mealanocytic
Mealanocytic. operation.
Lesions
Fuzzy logic Pattern classification
Thresholding-based Ensemble of dermoscopy 2 Classes:
Region of interest are
images: a Benign Melanocytic
extracted.
Renyi’s entropy perceptually uniform lesion and Melanoma
model
Otsu’s thresholding 5 Classes:
Segmentation and
Statistics Melanoma, Bullae, Region growing
Classification of skin
Seborrheic keratosis, segmentation is
lesions for disease
Adaptive thresholding Shingles and developed
diagnosis.
Squamous cell
Fuzzy logic
Iterative thresholding Goyal et al. (2019) highlight the significant impact of
Edge-based Edge detectors deep learning on medical imaging, with U-Net being a
powerful tool for skin lesion identification and segmentation.
Other methods Dynamic programming Bi et al. (2017) proposed multi-stage FCNs, improving
Hill-climbing algorithm accuracy through parallel integration strategies. Yuan et al.
(2017) introduced a 19-layer DCNN-based approach, using
the Jaccard Distance for loss function, tested with ISBI 2016
Banerjee et al. (2020) state that after pre-processing, and PH2 datasets. Goyal et al. (2017) demonstrated multi-
segmentation identifies the boundaries of the affected region. class segmentation algorithms using the 2017 ISBI challenge
Traditional methods like thresholding, area augmenting, and dataset. Vesal et al. (2018) and Goyal et al. (2019) presented
clustering are time-consuming and complex, making them a two-stage method with Faster-RCNN and a redefined U-
less effective for melanoma diagnosis. New approaches are Net. Soudani et al. (2019) applied deep learning models on
replacing these methods. Yacin Sikkandar et al. (2021) the ISIC-2017 dataset. Al-masni et al. (2018) used Fully
describe semi-automated image segmentation using a graph Convolutional Networks (FrCNs) for pixel feature learning in
where each pixel is a node, with additional nodes for dermoscopic images, achieving a Jaccard Index of 77.11%
background and foreground links. The graph is segmented on the ISIC-2017 test set. Table III compares segmentation
using Min-Cut/Max-Flow and Grabcut models, with the models and methods.
Gaussian Mixture Model (GMM) providing area details.
Tong et al. (2021) mention traditional segmentation methods V. PARAMETER TUNING METHODS FOR SKIN IMAGE
like threshold gradient vector flow, region growth, and CLASSIFICATION
morphology-based techniques, while CNN-based algorithms,
Banerjee et al. (2020) conducted a two-phase evaluation
like Full convolutional residual networks (FcRN) and U-Net
of methods to locate skin lesions using YOLOv3. The
networks, enhance segmentation by integrating multi-scale
model's ability to identify lesion locations in training images
contextual information and addressing challenges such as
was assessed using the IOU metric, claiming locations with
fixed receptive fields and gradient disappearance.
an IOU score above 80%. The second phase evaluated
sensitivity (Sen), specificity (Spe), the dice coefficient (Dic),
Generative adversarial networks (GANs) improve and the JACARD index (Acc). Sensitivity measured how
cutaneous lesion segmentation. Bi et al. (2017) proposed a well lesions were segmented, specificity quantified
multi-stage Fully Convolutional Network (FCN) with both segmented lesions, the dice coefficient examined
low-level and high-level semantic pyramids. Hasan et al. relationships with ground truth masks, and accuracy showed
(2020) introduced depth-wise separable convolution in the
total pixel-wise segmentation performance. For the methods for balancing multi-class classification datasets.
assessment metrics listed above, the formula is as follows: Bhardwaj et al. (2021) noted that advances in image
processing and machine learning have simplified feature
IOU = Area of Overlap/Area of Union extraction, inspired by Gabor and HSV filters, with
Sen = TP/TP+FN significant GPU and dataset advancements.
Spe = TN/TN+FP
Barata et al. (2013) proposed an approach for melanoma
Dic = 2xTP/(2xTP)+FP+FN identification using global and local characteristics,
Jac = TP/TP+FN+FP extracting texture, shape, and color features through
Acc = TP+TN/TP+TN+FN+FP segmentation and histograms. Esteva et al. (2017) used
True positive (TP), true negative (TN), false positive GoogleNet to classify over 100,000 skin lesion datasets.
(FP), and false negative (FN) represent the outcomes in Zhang et al. (2018) employed deep learning to classify four
image segmentation. TP pixels are accurately recognized and skin disorders using dermatoscopic images, achieving 86%
segmented, whereas FN pixels are not. Non-lesion pixels are accuracy on 1,067 images, highlighting the importance of
TN if correctly identified as non-lesion, while FP pixels are domain expertise and hierarchical structure. Table 5
inaccurately recognized as lesion pixels. Deep learning and represents the qualitative comparison models with well-
machine learning methods are employed for image known models used for classification.
classification. Table 4 presents the image classification
techniques used by these methods. TABLE V. QUALITATIVE COMPARISON MODELS WITH WELL KNOWN
MODELS USED FOR CLASSIFICATION

TABLE IV. IMAGE CLASSIFICATION USED BY DEEP LEARNING AND Classifiers Classes Title
MACHINE LEARNING METHODS 5 Classes:
Segmentation and
KNN, SVM, and Squamous cell,
Author & Classification of skin
Objective Method combined KNN and Shingles, Seborrheic
Year lesions for disease
SVM classifiers keratosis, Bullae,
Lee et al. Using pre-trained data and a Fine-tuned Neural diagnosis.
Melanoma
2018 Fine-Tune Neural Network, Pattern classification
it is easy to create and update AdaBoostMC 2 Classes:
of dermoscopy
a new challenge. Adaptive Boosting Melanoma and
images: a
By storing the data across the Multi channel Benign Melanocytic
perceptually uniform
network nodes, Artificial classification lesion
model
Neural Networks are 3 Classes: Automated
Noord et al. Artificial Neural
effective at recognizing non- KNN classifer Melanoma, Dysplatic Melanoma
2017 Networks
linear relationships between lesion and Benign Recognisation.
the dependent and
A methodological
independent parameters. 2 classes:
NN, SVM, 2-KNN, approach to the
Models of Convolutional Benign and
and Decision Trees classification of
Neural Networks are Melanoma
dermoscopy images
effective in selecting crucial
Computer Aided
Naranjo- characteristics automatically. 2 classes:
Convolutional Neural Statistical learning Diagnosis of
Torres et al. Instead of keeping the Melanoma and
Networks SVM Mealanocytic
2020 network node’s training data Mealanocytic
Lesions
in auxiliary memory, the
CNN model uses multi-layer
perceptrons to store it.
Structured and unstructured VII. CONCLUSION
data may be processed using The skin cancer, including melanoma, BCC, and SCC,
Deep Neural Networks-
Yadav et al.
based algorithms.
Deep Neural poses a significant global health challenge. Differentiating
2019 Networks types of skin cancer through clinical signs, risk factors, and
Unlabeled data may still be
used by the models and molecular markers is crucial. New diagnostic tests such as
produce a better result. molecular tests, dermoscopy, and histopathology have
improved early diagnosis and patient outcomes. The dynamic
field of skin cancer treatment includes targeted drugs,
VI. SKIN LESION CLASSIFICATION radiation therapy, immunotherapy, and excision procedures,
Researchers (Lima et al. 2021; Hameed et al. 2018) offering hope for complex cases like metastasized melanoma.
evaluated 1,011 images using a 7-point system, creating Genetic and molecular screening enable tailored treatment
multimodal CNNs that process clinical and dermoscopic approaches. Prevention through community campaigns and
images, along with metadata. They used two Inception v3 sun protection is vital. Eradicating this pervasive disease
CNNs, each representing one of seven criteria, trained to requires enhancing patient outcomes through
detect abnormalities. The ISIC2018 (HAM10000) collection interdisciplinary approaches, individualized treatments, and
included 10,015 dermoscopic images for identifying benign, preventive measures. Cross-disciplinary research and
necrotic, actinic keratosis, or melanoma lesions, using two collaboration are essential for effective skin cancer
pre-trained Inception-v3 CNNs for combined output. management.
REFERENCES
Gessert et al. (2020) analyzed the HAM10000 training
set, 7-point, and ISIC archive (2018) datasets, testing Multi- [1] Daghrir, J., Tlig, L., Bouchouicha, M., & Sayadi, M. (2020,
Crop, Recurrent output modules, and Patch-based Attention September). Melanoma skin cancer detection using deep learning and
classical machine learning techniques: A hybrid approach. In 2020 5th
against baseline deep-feature extractors like Inception-v3,
DenseNet-121, and SEResnext50. They explored four
 international conference on advanced technologies for signal and [22] Diame, ZE, Al-Berry, MN, Salem, MAM & Roushdy, M 2021, ‘Deep
image processing (ATSIP) (pp. 1-5). IEEE. Learning Architiectures for Aided Melanoma Skin Disease
[2] 2. Kadampur, M. A., & Al Riyaee, S. (2020). Skin cancer Recognition: A Review’, Proceedings of theInternational Mobile,
detection: Applying a deep learning based model driven architecture Intelligent, and Ubiquitous Computing Conference (MIUCC) pp. 324-
in the cloud for classifying dermal cell images. Informatics in 329.
Medicine Unlocked, 18, 100282.I. S. Jacobs and C. P. Bean, “Fine [23] Rojaramani, D 2020, ‘An analysis for melanoma diagnosis in
particles, thin films and exchange anisotropy,” in Magnetism, vol. III, dermascopic images using integrated segmentation and multiple
G. T. Rado and H. Suhl, Eds. New York: Academic, 1963, pp. 271– feature extraction’.
350. [24] Mishra, VK, Kumar, S & Shukla, N 2017, ‘Image acquisition and
[3] Berhil, S., Benlahmar, H., & Labani, N. (2020). A review paper on techniques to perform image acquisition’, SAMRIDDHI: A Journal of
artificial intelligence at the service of human resources management. Physical Sciences, Engineering and Technology, vol. 9(01), pp.21-24.
Indonesian Journal of Electrical Engineering and Computer Science, [25] Tumpa, PP & Kabir, MA 2021, ‘An artificial neural network based
18(1), 32-40. detection and classification of melanoma skin cancer using hybrid
[4] Kliegr, T., Bahník, Š., & Fürnkranz, J. (2021). A review of possible texture features’, Sensors International, vol. 2, pp.100128.
effects of cognitive biases on interpretation of rule-based machine [26] Oliveira, RB, Mercedes Filho, E, Ma, Z, Papa, JP, Pereira, AS&
learning models. Artificial Intelligence, 295, 103458. Tavares, JMR 2016, ‘Computational methods for the image
[5] Hajiarbabi, M. (2023). Skin cancer detection using multi-scale deep segmentation of pigmented skin lesions: a review’, Computer
learning and transfer learning. Journal of Medical Artificial methods and programs in biomedicine, vol. 131, pp.127-141.
Intelligence, 6. [27] Banerjee, S, Singh, SK, Chakraborty, A, Das, A & Bag, R 2020,
[6] Lu, X., & Firoozeh Abolhasani Zadeh, Y. A. (2022). Deep Learning‐ ‘Melanoma diagnosis using deep learning and fuzzy logic’,
Based Classification for Melanoma Detection Using XceptionNet. Diagnostics, vol. 10(8), pp.577.
Journal of Healthcare Engineering, 2022(1), 2196096. [28] Yacin Sikkandar, M, Alrasheadi, BA, Prakash, NB, Hemalakshmi,
[7] Khanzode, K. C. A., & Sarode, R. D. (2020). Advantages and GR, Mohanarathinam, A & Shankar, K 2021, ‘Deep learning based an
disadvantages of artificial intelligence and machine learning: A automated skin lesion segmentation and intelligent classification
literature review. International Journal of Library & Information model’, Journal of ambient intelligence and humanized computing,
Science (IJLIS), 9(1), 3. vol. 12(3), pp.3245-3255.
[8] Ullah, Z., Al-Turjman, F., Mostarda, L., & Gagliardi, R. (2020). [29] Szegedy, C, Ioffe, S, Vanhoucke, V & Alemi, AA 2017, ‘Inception-
Applications of artificial intelligence and machine learning in smart v4, inception-resnet and the impact of residual connections on
cities. Computer Communications, 154, 313-323. learning’, In Thirty-first AAAI conference on artificial intelligence.
[9] Hermosilla, P., Soto, R., Vega, E., Suazo, C., & Ponce, J. (2024). Skin [30] Tong, X, Wei, J, Sun, B, Su, S, Zuo, Z & Wu, P 2021, ‘ASCU-Net:
Cancer Detection and Classification Using Neural Network attention gate, spatial and channel attention u-net for skin lesion
Algorithms: A Systematic Review. Diagnostics, 14(4), 454. segmentation’, Diagnostics, vol. 11(3), pp.501.
[10] Lemaignan, S., Warnier, M., Sisbot, E. A., Clodic, A., & Alami, R. [31] Yu, Z, Jiang, X, Zhou, F, Qin, J, Ni, D, Chen, S, Lei, B & Wang, T
(2017). Artificial cognition for social human–robot interaction: An 2018,‘Melanoma recognition in dermoscopy images via aggregated
implementation. Artificial Intelligence, 247, 45-69. deep convolutional features’, IEEE Transactions on Biomedical
[11] Wu, Y., Chen, B., Zeng, A., Pan, D., Wang, R., & Zhao, S. (2022). Engineering, vol. 66(4), pp.1006-1016.
Skin cancer classification with deep learning: a systematic review. [32] Song, L, Lin, J, Wang, ZJ & Wang, H 2019, ‘Dense-residual attention
Frontiers in Oncology, 12, 893972. network for skin lesion segmentation’, Proceedings of the
[12] Thrall, J. H., Li, X., Li, Q., Cruz, C., Do, S., Dreyer, K., & Brink, J. International Workshop on Machine Learning in Medical
(2018). Artificial intelligence and machine learning in radiology: Imaging,Springer, Cham, pp. 319-327.
opportunities, challenges, pitfalls, and criteria for success. Journal of [33] Bi, L, Kim, J, Ahn, E, Kumar, A, Fulham, M & Feng, D 2017,
the American College of Radiology, 15(3), 504-508. ‘Dermoscopic image segmentation via multistage fully convolutional
[13] Dildar, M., Akram, S., Irfan, M., Khan, H. U., Ramzan, M., networks’, IEEE Transactions on Biomedical Engineering, vol. 64(9),
Mahmood, A. R., ... & Mahnashi, M. H. (2021). Skin cancer pp.2065-2074.
detection: a review using deep learning techniques. International [34] Hasan, MK, Dahal, L, Samarakoon, PN, Tushar, FI& Martí, R
journal of environmental research and public health, 18(10), 5479. 2020,‘DSNet: Automatic dermoscopic skin lesion segmentation’,
[14] Connor, C. W. (2019). Artificial intelligence and machine learning in Computers in Biology and Medicine, vol. 120, p.103738.
anesthesiology. Anesthesiology, 131(6), 1346-1359. [35] Adegun, AA & Viriri, S 2020, ‘FCN-based DenseNet framework for
[15] Wysocki, O., Davies, J. K., Vigo, M., Armstrong, A. C., Landers, D., automated detection and classification of skin lesions in dermoscopy
Lee, R., & Freitas, A. (2023). Assessing the communication gap images’, IEEE Access, vol. 8, pp.150377-150396.
between AI models and healthcare professionals: Explainability, [36] Ashour, AS, Nagieb, RM, El-Khobby, HA, Abd Elnaby, MM & Dey,
utility and trust in AI-driven clinical decision-making. Artificial N 2021, ‘Genetic algorithm-based initial contour optimization for skin
Intelligence, 316, 103839. lesion border detection’, Multimedia Tools and Applications, vol.
[16] Rupali, M., & Amit, P. (2017). A review paper on general concepts of 80(2), pp.2583-2597.
artificial intelligence and machine learning. International Advanced [37] K. Ding Azad, M, Bozorg-Haddad, O & Chu, X 2018, ‘Flower
Research Journal in Science, Engineering and Technology, 4(4), 79- pollination algorithm (FPA)’, In Advanced optimization by nature-
82. inspired algorithms pp. 59-67. Springer, Singapore.
[17] Collenette, J., Atkinson, K., & Bench-Capon, T. (2023). Explainable [38] Isah, OR, Usman, AD & Tekanyi, AM 2017, ‘Medical image
AI tools for legal reasoning about cases: A study on the European segmentation through bat-active contour algorithm’.
Court of Human Rights. Artificial Intelligence, 317, 103861. [39] Goyal, M, Oakley, A, Bansal, P, Dancey, D & Yap, M.H, 2019, ‘Skin
[18] Winkler, D. A. (2020). Role of artificial intelligence and machine lesion segmentation in dermoscopic images with ensemble deep
learning in nanosafety. Small, 16(36), 2001883. learning methods’, IEEE Access, 8, pp.4171-4181.
[19] Dildar, M, Akram, S, Irfan, M, Khan, H.U, Ramzan, M, Mahmood, [40] Goyal, M, Yap, MH & Hassanpour, S 2017, ‘Multi-class semantic
AR, Alsaiari, SA, Saeed, AHM, Alraddadi, MO & Mahnashi, MH segmentation of skin lesions via fully convolutional networks’, arXiv
2021, ‘Skin cancer detection: a review using deep learning preprint arXiv:1711.10449.
techniques’, International journal of environmental research and [41] Vesal, S, Malakarjun Patil, S, Ravikumar, N & Maier, AK 2018, ‘A
public health, vol. 18(10), pp.5479. multi-task framework for skin lesion detection and segmentation’, In
[20] Schierbeck, J, Vestergaard, T & Bygum, A 2019,‘Skin cancer OR 2.0 Context-Aware Operating Theaters, Computer Assisted
associated genodermatoses: a literature review’, Acta Derm Venereol, Robotic Endoscopy, Clinical Image-Based Procedures, and Skin
vol. 99(4), pp.360-369. Image Analysis pp. 285-293. Springer, Cham.
[21] Shorfuzzaman, M 2021, ‘An explainable stacked ensemble of deep [42] Al-masni Wang, D, Pang, N, Wang, Y & Zhao, H 2021, ‘Unlabeled
learning models for improved melanoma skin cancer detection’, skin lesion classification by self-supervised topology clustering
Multimedia Systems, pp.1-15. network’, Biomedical Signal Processing and Control, vol. 66,
pp.102428.
[43] Y, Chao, M & Lo, YC 2017, ‘Automatic skin lesion segmentation binary classification support’, Biomedical Signal Processing and
using deep fully convolutional networks with jaccard distance’, IEEE Control, vol. 62, pp.102041.
transactions on medical imaging, vol.36(9), pp.1876-1886. [50] Gessert, N, Sentker, T, Madesta, F, Schmitz, R, Kniep, H, Baltruschat, I,
[44] Lee, YC, Jung, SH & Won, HH 2018, ‘WonDerM: Skin lesion Werner, R & Schlaefer, A 2019, ‘Skin lesion classification using cnns
classification with fine-tuned neural networks’, arXiv preprint with patch-based attention and diagnosis-guided loss weighting’,
arXiv:1808.03426. IEEE Transactions on Biomedical Engineering, vol. 67(2), pp.495-
[45] Van Noord, N & Postma, E 2017, ‘Learning scale-variant and scale- 503.
invariant features for deep image classification’, Pattern Recognition, [51] Bhardwaj, A & Rege, PP 2021,‘Skin lesion classification using deep
vol. 61, pp.583-592. learning’, Proceedings of the Advances in Signal and Data Processing,
[46] Yadav, SS & Jadhav, SM 2019, ‘Deep convolutional neural network pp. 575-589. Springer, Singapore.
based medical image classification for disease diagnosis’, Journal of [52] Esteva, A, Kuprel, B, Novoa, RA, Ko, J, Swetter, SM, Blau, HM &
Big Data, vol. 6(1), pp.1-18. Thrun, S 2017, ‘Dermatologist-level classification of skin cancer with
[47] Naranjo-Torres, J, Mora, M, Hernández-García, R, Barrientos, RJ, deep neural networks’, nature, vol. 542(7639), pp.115-118.
Fredes, C & Valenzuela, A 2020,‘A review of convolutional neural [53] Zhang, X, Wang, S, Liu, J & Tao, C 2018, ‘Towards improving
network applied to fruit image processing’, Applied Sciences, vol. diagnosis of skin diseases by combining deep neural network and
10(10), pp.3443. human knowledge’, BMC medical informatics and decision making,
[48] Lima, DM, Rodrigues-Jr, JF, Brandoli, B, Goeuriot, L & Amer-Yahia, vol. 18(2), pp.69-76.
S 2021, ‘Dermadl: advanced convolutional neural networks for
computer-aided skin-lesion classification’, SN Computer Science,
vol. 2(4), pp.1-13.
[49] Harangi, B, Baran, A&Hajdu, A 2020,‘Assisted deep learning
framework for multi-class skin lesion classification considering a